Iris-Z

Appearance and geometry, one model

Iris-Z is a general image generator that produces appearance and geometry together. One set of weights, with no depth head and no second network. It is the first general-purpose generator to beat state-of-the-art task-specific models like MoGe-3.

Paper Code Weights Demo

Alejandro Cañada^†Chema GarabitoSperidlabs · † project lead

Gallery

Image + Depthboth generated from a prompt

Image → Depthdepth generated from an image

Depth → Imageimage generated from a depth map

One model, every task

Iris-Z: the prompt, the image and the depth are encoded into one token sequence and processed by a single diffusion transformer

One sequence, one backbone

Iris-Z comes from a pretrained text-to-image diffusion transformer. There is no depth head and no second network: depth is stored as normalised log ray distance, encoded as an image by the same autoencoder and placed in the same token sequence, so image and depth condition each other at every step of denoising.

Scale

The larger backbone is the more accurate, and the slower. A specialist is still much cheaper per image.

Mean AbsRel × 100 over each benchmark table, and the time of one prediction per image at 512².1 × H200

Metric	MoGe-3	Iris-Z-small	Iris-Z-big
Marigold’s sets ↓	3.12	3.11	2.92
MoGe-3’s sets ↓	3.70	3.84	3.56
Time ↓	0.4 s	4.8 s	15.0 s

Model	Marigold’s sets ↓	MoGe-3’s sets ↓	Time ↓
Model	Marigold’s sets ↓	MoGe-3’s sets ↓	Time ↓
MoGe-3ViT-G, 1 pass	3.12	3.70	0.4 s
Iris-Z-small9B, 40 passes	3.11	3.84	4.8 s
Iris-Z-big32B, 40 passes	2.92	3.56	15.0 s

Evaluations and benchmarks

AbsRel × 100 on seven of MoGe-3’s benchmarks, relative-error L1 alignment. Bold is the best in each column.MoGe-3’s protocol

Metric	MoGe-3	Marigold V2	Iris-Z-small	Iris-Z-big
NYUv2 ↓	2.63	3.34	2.40	2.28
KITTI ↓	3.41	4.96	3.53	2.98
ETH3D ↓	2.01	4.83	2.18	1.87
iBims-1 ↓	2.07	2.81	2.20	1.99
Sintel ↓	10.85	14.88	11.50	10.67
HAMMER ↓	2.30	2.80	2.22	1.93
DIODE ↓	2.65	4.18	2.88	3.18
Mean ↓	3.70	5.40	3.84	3.56

Model	NYUv2 ↓	KITTI ↓	ETH3D ↓	iBims-1 ↓	Sintel ↓	HAMMER ↓	DIODE ↓	Mean ↓
Model	NYUv2 ↓	KITTI ↓	ETH3D ↓	iBims-1 ↓	Sintel ↓	HAMMER ↓	DIODE ↓	Mean ↓
MoGe-3	2.63	3.41	2.01	2.07	10.85	2.30	2.65	3.70
Marigold V2	3.34	4.96	4.83	2.81	14.88	2.80	4.18	5.40
Iris-Z-small	2.40	3.53	2.18	2.20	11.50	2.22	2.88	3.84
Iris-Z-big	2.28	2.98	1.87	1.99	10.67	1.93	3.18	3.56

AbsRel × 100 on Marigold’s benchmarks, RANSAC alignment in log depth. Bold is the best in each column.Marigold’s protocol

Metric	MoGe-3	Marigold V2	Iris-Z-small	Iris-Z-big
NYUv2 ↓	2.65	3.58	2.07	1.90
ETH3D ↓	2.06	2.78	2.12	1.76
ScanNet ↓	2.86	3.65	2.56	2.25
KITTI ↓	3.98	5.41	4.58	4.16
DIODE ↓	4.03	5.07	4.22	4.50
Mean ↓	3.12	4.10	3.11	2.92

Model	NYUv2 ↓	ETH3D ↓	ScanNet ↓	KITTI ↓	DIODE ↓	Mean ↓
Model	NYUv2 ↓	ETH3D ↓	ScanNet ↓	KITTI ↓	DIODE ↓	Mean ↓
MoGe-3	2.65	2.06	2.86	3.98	4.03	3.12
Marigold V2	3.58	2.78	3.65	5.41	5.07	4.10
Iris-Z-small	2.07	2.12	2.56	4.58	4.22	3.11
Iris-Z-big	1.90	1.76	2.25	4.16	4.50	2.92

Cite this work

@techreport{canada2026irisz,
  title   = {Iris-Z: Scaling One Generative Model Past
             Specialist Geometry Models},
  author  = {Ca{\~n}ada, Alejandro and Garabito, Chema},
  institution = {Speridlabs},
  year    = {2026}
}